A self-experiment that provokes reflection.

2026-05-21 · Author: Beni Gafner The question sounds like pub talk, yet it has haunted me since I saw my own face in a video a few days ago – a video that was never recorded. At first, I was a drawn character, angular and friendly, almost like something from an animated series. Seconds later, I was cycling – photorealistic, with my build, my eyes, my way of looking – through a race track as a cyclist, crashed unceremoniously, picked myself up, and nevertheless crossed the finish line first. All invented. All generated. All in a quality that unsettled me.

From tool to reality machine

As a journalist, I have worked with images for as long as I can remember. I have learned to distrust them, to check sources, to recognise cuts, to question manipulations. What I witnessed in this self-experiment overrides these learned reflexes. It was no clumsy fake, no obvious montage. It was a human being – me – in a situation that never took place. And no one who knows me would have suspected anything upon first glance. Your browser does not support embedded video. The technology behind this is no longer secret knowledge. Open-access models can now generate videos from just a few seconds of reference material that deceive the eye. What until recently occupied entire VFX studios can now run on a powerful computer – or directly in the browser, for a small subscription fee.

The dangers are no longer theoretical

The loss of trust in the image. For decades, moving images were considered reliable witnesses. This self-evident truth is eroding. If every video could be real – or not – the medium loses its evidentiary power. The legal consequences are foreseeable: recordings will become challengeable in court, even if they are authentic. The reversal of the burden of proof. Just as problematic is the effect researchers call the “liar’s dividend”: anyone caught doing something unpleasant on real video footage will be able to claim it is a fake in the future. Doubt alone is enough. What does not need to be disproven does not need to be admitted. Real-time political manipulation. A forged statement from a Bundesrätin, 48 hours before a vote, spread via social networks – who can correct that fast enough? Switzerland, with its frequent referendums, is particularly vulnerable. A Hollywood production is no longer needed to overturn a campaign. All that is required is a laptop and an afternoon. Identity theft in everyday life. My face on a bicycle that does not belong to me, in a race that does not exist – that is the harmless version. The unpleasant one: my face in an advertisement I never booked. In a political message that is not mine. In material that I do not wish to write about here, but which exists and primarily affects women. The erosion of journalism. If all visual material is broadly regarded as manipulable, those recordings that truly document something – war crimes, police violence, corruption – also lose their impact. Research becomes more expensive, slower, riskier. The gap between reputable newsrooms and the speed of disinformation continues to widen. Private harm. Bullying in schools is already functioning today with fabricated images of classmates. What happens when every fourteen-year-old has a tool at their disposal that allows classmates to be inserted into any scene? We must discuss this before the first suicides make headlines.

What frightened me most about my own experiment

It was not the quality of the material. It was how quickly I became accustomed to the sight. After three or four loops, the video almost felt like a real memory. If my own brain, which should know better, can be so easily deceived – how will it affect readers seeing the material for the first time, between two push notifications, on their way home on a tram?

What is now necessary I am not a fan of hasty bans. The technology has its legitimate applications: in film production, in education, in accessibility. But we need three things, and we need them soon: First, a mandatory labelling requirement for synthetically generated material, with teeth – not as a non-binding recommendation. Second, technical watermarks on the producer side, built into the models themselves, ideally coordinated internationally. Third, and most importantly: media literacy as a compulsory subject, from primary school age. Those growing up with this technology must learn how to confront it early on.

The opening question remains

Are the capabilities of AI dangerous? After this self-experiment, I say: the capabilities themselves are not. What is dangerous is the gap between what the technology can do and what our institutions, laws, and reflexes are prepared for. This gap is growing faster than we are closing it. By the way, my avatar is still crossing the finish line. Having stood up after the crash, smiling, undeterred. Reality was never its concern. Ours must remain so.

Questions to uncensored AI: https://podcasts.politiq.ch/episoden/ep-2-wer-kontrolliert-die-ki/

Book recommendation: https://beat-w-meier.ch/buecher/der-goldene-kaefig/